Governance Framework · v1.0

Claresia SCUDO

The AI shield for the EU enterprise.

Five pillars — Scansione, Catena di custodia, Uso conforme, Dati controllati, Operatori verificati — keeping every Claresia AI agent inside policy, EU AI Act, and Garante boundaries. Italian-first. Cryptographically verifiable. On by default.

Why we built this

A named answer to "how do you keep your agents from going off the rails?"

AI risk is real, immediate, and asymmetric. A single autonomous skill that emails the wrong recipient, triggers a workflow against the wrong tenant, or leaks a codice fiscale into a third-party model can unwind months of trust — and trigger Garante or NIS2 reporting obligations within hours.

Claresia SCUDO is our answer. Five pillars, cryptographically anchored, EU AI Act + Garante explicit, on by default for every tenant. It maps one-to-one with Glean's AWARE framework so head-to-head security reviews are honest. It extends AWARE with a fifth EU-specific pillar — Uso conforme — that auto-stamps Article 50 disclosures and regenerates the Annex IV technical-doc pack every night.

Scudo is the Italian word for shield. It drops cleanly into a CISO slide. It doesn't oversell. It is the explicit aggregation of governance contracts the platform already enforces.

The 5 pillars

SCUDO in detail

Every pillar declares: what it is, the control it provides, the implementation in Claresia today, how it maps to EU regulation, and what the customer sees on Trust Center.

S

Scansione — Pre-execution Scan

AWARE — Autonomous guardrails + Restricted topics

Every write-class action passes through a per-tenant restricted-topics policy + alignment model before any token leaves Claresia infrastructure.

What it is

Per-tenant restricted_topics.json (Italian defaults: codice fiscale exposure, biometric inference, fully-automated employment decisions per Garante 2024) + a Haiku-grade alignment model that re-reads the planned tool call against the tenant rubric and flags concerns. severity≥block ⇒ rejected.

The control it provides

No write-class action ever fires without a fresh policy + alignment check bound to the tenant rubric.

How it's implemented
  • lib/governance/restricted-topics.ts — policy engine, per-tenant rules
  • lib/governance/alignment-prescan.ts — async pre-execution check
  • Wired at cc-073 LLM Gateway entry + cc-061 Roster Engine action dispatcher
  • Italian-default policy shipped at site/src/data/restricted-topics-italy-default.json
EU regulatory mapping
  • EU AI Act Art. 9 — Risk management system
  • EU AI Act Art. 14 — Human oversight
  • GDPR Art. 22 — Restriction on solely-automated decisions
  • Garante Provv. n. 9978728/2024 (riconoscimento facciale)
  • Statuto Lavoratori Art. 4 (monitoraggio indiretto)
  • ISO/IEC 42001 §8.3 Operation
What the customer sees

Trust Center › Governance › Restricted-topics editor (per-tenant) + Audit log entries showing pre-scan verdict per skill invocation.

C

Catena di custodia — Cryptographic Audit Chain

AWARE — Ecosystem observability

Every privileged action emits a SHA-256 hash-chained governance event. Tamper-evident, customer-verifiable, 7-year retention by default.

What it is

Each event carries the SHA-256 of the previous event for the same (tenant_id, stream), forming a Merkle-style append-only ledger. Anchors into cc-050 Hub. cc-064 Telemetry forwards records to the customer per-tenant SIEM destination.

The control it provides

Complete, tamper-evident, cryptographically-verifiable history of every AI action with deterministic offline replay.

How it's implemented
  • lib/governance/audit-chain.ts — chain emitter + verifier
  • cc-050 Intelligence Hub governance_events table
  • cc-064 Telemetry per-tenant SIEM streaming destination
  • verify-chain CLI shipped under MIT license
EU regulatory mapping
  • EU AI Act Art. 12 — Record-keeping
  • EU AI Act Art. 19 — Logs to be kept by deployer
  • GDPR Art. 30 — Records of processing activities
  • NIS2 Art. 21(2)(c) + 23 — Incident handling
  • Italian D.Lgs. 138/2024 ACN/CSIRT incident notification
  • DORA Art. 11–12 — ICT incident classification
  • ISO/IEC 42001 §9.1 Monitoring + measurement
What the customer sees

Trust Center › Audit › Downloadable signed audit log (JSON) + per-tenant SIEM streaming destination + standalone verify-chain CLI.

U

Uso conforme — EU AI Act + Garante Alignment

(no direct AWARE pillar — SCUDO EU-specific differentiator)

Every output meeting Article 50 / Article 22 / Annex III triggers is auto-stamped with locale-aware disclosure. Annex IV technical-doc pack regenerates nightly per tenant.

What it is

Outputs facing natural persons or matching Annex III categories carry an Article 50 transparency stamp in the user locale (it/en at minimum). Article 22 outputs are reserved-to-human. Annex IV pack auto-built from deployed skill IRs and exported per tenant.

The control it provides

Customer cannot ship a non-compliant AI output by accident. Disclosure is automatic; technical documentation is always export-ready.

How it's implemented
  • lib/governance/eu-ai-act-disclosure.ts — locale-aware Article 50 stamp
  • uso-conforme-classifier.ts (helper) — Annex III labelling at deploy time
  • Auto-generated Annex IV pack at governance.claresia.com/tenants/{tid}/annex-iv.pdf, regenerated nightly
  • Locale tables: it (Italian-first) + en (fallback) + customer-extensible
EU regulatory mapping
  • EU AI Act Art. 50 — Transparency obligations
  • EU AI Act Art. 11 + Annex IV — Technical documentation
  • EU AI Act Art. 13 — Information for deployers
  • EU AI Act Art. 53 — GPAI provider obligations
  • GDPR Art. 22 — Reservation of human decision
  • Garante Provv. ChatGPT 2023 — Transparency
  • Codice della Privacy Art. 122 — Italian consent posture
What the customer sees

Trust Center › EU AI Act conformance pack (per-tenant Annex IV PDF + SHA-256 checksum) + every AI surface stamps disclosures in EN/IT.

D

Dati controllati — Egress Control / LLM Gateway

AWARE — Work context + Glean Protect data-loss controls

Every LLM call from every Claresia function routes through cc-073 LLM Gateway: model allowlist, EU region pinning, bidirectional PII redaction, per-tenant workspace, hard cost cap.

What it is

Single egress path. Default region eu-south-1 (Milano). Per-tenant Anthropic / OpenAI workspace via admin APIs. Microsoft Presidio + custom NER bidirectional redaction. Hard quota. Full audit log streamed to customer SIEM. Optional BYOC mode (gateway in customer VPC, customer-held keys).

The control it provides

Customer data flows out of Claresia perimeter only through one auditable choke point with explicit allowlisted destinations.

How it's implemented
  • lib/governance/llm-gateway.ts — reference gateway client
  • Microsoft Presidio (Python sidecar) for PII; Lakera/Prompt Armor optional overlays
  • Anthropic Enterprise admin-API + OpenAI Enterprise admin-API for per-tenant workspaces
  • cc-073 is the canonical roadmap item; this framework defines its contract surface
EU regulatory mapping
  • GDPR Art. 5(1)(c) — Data minimisation
  • GDPR Art. 28 — Sub-processor transparency
  • GDPR Art. 32 — Security of processing
  • GDPR Art. 44–49 — International transfers (Schrems II)
  • EU AI Act Art. 10 — Data and data governance
  • NIS2 Art. 21(2)(d) — Supply-chain security
  • Garante Provv. trasferimenti USA — SCC + TIA
  • ISO/IEC 27001 A.5.34 PII protection
What the customer sees

Trust Center › Sub-processors page (auto-populated from gateway log) + per-tenant Egress dashboard (model usage by region + redaction-rate counters).

O

Operatori verificati — Identity + Intent Verification

AWARE — Actor intent

Every action is bound to a verified human (WorkOS / SCIM) + an immutable skill IR hash + a declared intent. No anonymous actions, no unapproved skill versions.

What it is

WorkOS / SSO / SCIM-provisioned identity, MFA via customer IdP. Each skill carries an immutable ir_hash; only Roster-approved IR hashes can dispatch. For autonomous schedules, the human owner of the schedule is captured + the tenant policy permitting unattended execution.

The control it provides

No action is anonymous; no action runs against an unapproved skill version; every parameter preserved for forensics.

How it's implemented
  • WorkOS as canonical IdP layer (cc-059 admin console)
  • SCIM 2.0 for deprovisioning (immediate on IdP removal)
  • Skill IR contract (cc-060) ships immutable ir_hash + restricted_topics field
  • Roster Engine (cc-061) refuses dispatch of any IR hash not in tenant approved-rosters
  • governance_event.actor_* fields hold employee_id + ir_hash on every record
EU regulatory mapping
  • EU AI Act Art. 14 — Human oversight (named person)
  • EU AI Act Art. 26 — Deployer obligations
  • GDPR Art. 5(1)(f) — Integrity & confidentiality
  • NIS2 Art. 21(2)(i) — Multi-factor authentication
  • Statuto Lavoratori Art. 4 — RSU/RSA + cohort-suppression mode
  • Garante Provv. n. 232/2024 (proporzionalità)
  • ISO/IEC 27001 A.5.15–A.5.18 Access control family
What the customer sees

Admin console › Roster (approved-skills view) + Audit log filterable by employee + Telemetry-suppression toggle (cohort-only mode for Statuto compliance).

Architecture

How SCUDO intersects the Claresia platform

SCUDO is not a layer bolted on top — it is the governance contract surface every layer of the platform already enforces. Below: the six layers and the SCUDO pillars that gate each transition.

Layer 6  End-user surfaces  (Teams · Slack · Adaptive Cards · Browser ext)
                    ↓ all output stamped under [U] Uso conforme
Layer 5  Distribution adapters  (cc-063 Claude · cc-065 Copilot · cc-070 ChatGPT)
                    ↓ dispatches gated by [O] Operatori verificati
Layer 4  Hub UX + Roster  (cc-060 Hub UX · cc-061 Roster Engine)
                    ↓ Roster enforces approved IR hashes
Layer 3  Function execution  (Sailford · Forge · Boss · Ledger · Gatespic · Takecare · Steve · Clawshield · Zottos)
                    ↓ every LLM call → [D] Dati controllati Gateway
                    ↓ every write-class action → [S] Scansione
Layer 2  cc-050 Intelligence Hub  (provenance · tenants · rosters)
                    ↓ holds the [C] Catena di custodia
Layer 1  Tenant infra  (Postgres EU · WorkOS · storage · SIEM destination)
      
Default deny

Skill IRs labelled risk_class: high are default-denied until the tenant admin explicitly approves them into Roster.

Two-person oversight

High-risk activation requires admin approval in cc-059 and a checked-in Article 14 human-oversight policy file.

Italian-first defaults

Restricted-topics, locale, region all default to Italy + EU. eu-south-1 Milano. Italian DPA. EN/IT stamps.

vs Glean AWARE

An honest comparison

Glean's AWARE is a serious framework. SCUDO maps one-for-one to make head-to-head reviews easy. Where Glean is honestly ahead today (broader connectors, named SOAR partner), we say so. Where SCUDO pulls ahead — explicit EU AI Act stamping, Garante-cited defaults, cryptographically-verifiable chain — we name it cleanly.

Feature
Glean
SCUDO
Note
Pre-execution alignment scan on write actions
=
=
Both ship; SCUDO ships Italian-default rules.
Restricted-topics policy editable per tenant
=
=
SCUDO ships Garante-cited defaults out of the box.
Real-time per-action risk scoring
=
=
Both compute + emit risk scores.
Audit log streamed to customer SIEM
=
=
Both ship; SCUDO additionally hash-chains the records.
Cryptographically-verifiable audit chain (Merkle-style)
SCUDO ships SHA-256 chain + offline verifyChain() CLI under MIT.
EU AI Act Article-50 runtime stamping
SCUDO auto-stamps in EN/IT; Glean handles disclosure in DPA copy.
Auto-generated Annex IV technical-doc pack per tenant
SCUDO regenerates nightly; surfaces as downloadable PDF.
Garante provvedimenti citation in restricted-topics
SCUDO ships per-rule citations (e.g. Garante 2024 n.9978728).
Italian-locale defaults out of the box (codice fiscale, IBAN-IT)
SCUDO ships Italy-specific NER + restricted-topics defaults.
Statuto dei Lavoratori art. 4 telemetry-suppression toggle
SCUDO ships cohort-only mode + RSU/RSA template.
Single LLM Gateway as only egress path (per-tenant model allowlist)
=
SCUDO defaults eu-south-1 Milano + bidirectional Presidio.
Named SOAR partner integration (Tines)
On cc-058 enterprise-readiness roadmap. Mitigate via SIEM streaming today.
DSPM partner integration (Palo Alto Cortex Cloud)
On cc-058 roadmap.
AI security partner co-development (Palo Alto Prisma AIRS)
No commercial AI-security partner today; Garante alignment is the differentiator.
100+ connectors with ACL inheritance
Connector v1 framework on cc-058 roadmap; first 20 connectors targeted Q3 2026.
ISO 42001 certified
SCUDO maps explicitly to §6.1/8.3/8.4/9.1; certification planned Q4 2026.
FedRAMP
Neither vendor pursues. SCUDO is Italy-first / EU-resident by design.

Sources: glean.com/blog/agentic-security-aware, glean.com/security.

Italian compliance map

EU AI Act · Garante · Statuto dei Lavoratori

The full pillar-by-pillar regulatory map lives in compliance-mapping-italy.md. Highlights below.

EU AI Act
  • Art. 9 — Risk management → [S]
  • Art. 10 — Data governance → [D]
  • Art. 11 + Allegato IV — Technical doc → [U]
  • Art. 12 — Record keeping → [C]
  • Art. 14 — Human oversight → [S][O]
  • Art. 22 GDPR — Solely-automated → [U]
  • Art. 50 — Transparency → [U]
  • Art. 53 — GPAI provider → [U]
Garante (Italian DPA)
  • Provv. ChatGPT 2023 — Transparency baseline
  • Provv. 2023 (videosorveglianza intelligente) — Workplace biometric
  • Provv. 31/2024 — AI in HR decision support
  • Provv. 232/2024 — Worker monitoring proportionality
  • Provv. 9978728/2024 — Riconoscimento facciale → blocked default

Restricted-topics ships citation-by-citation references. New Garante notices land in tenant policies within 14 days of publication.

Statuto Lavoratori art. 4

cc-064 telemetry per-employee logging is indirect monitoring under Italian law. SCUDO ships:

  • Italian DPA addendum naming the tool
  • Telemetry-suppression mode (cohort-only)
  • RSU/RSA agreement template under docs/it/templates/
  • Onboarding wizard prompt: "Hai ottenuto l'accordo RSU/RSA?"
GDPR
  • Art. 5(1)(c) — Minimisation → [D]
  • Art. 22 — Auto decisions → [U][S]
  • Art. 28 — Sub-processor → [D]
  • Art. 30 — RoPA → [C]
  • Art. 32 — Security → all pillars
  • Art. 33–34 — Breach notify → [C]
  • Art. 44–49 — Transfers → [D]
NIS2 + DORA
  • NIS2 Art. 21(2)(c)+(d)+(i) — Incident handling, supply chain, MFA
  • NIS2 Italia D.Lgs. 138/2024 — ACN/CSIRT notification (chain feeds)
  • DORA Art. 11–12 — ICT incident classification
ISO 27001 + 42001
  • ISO 27001 A.5.15–A.5.18 — Access control → [O]
  • ISO 27001 A.5.34 — PII protection → [D][S]
  • ISO 27001 A.8.15 — Logging → [C]
  • ISO 42001 §6.1 + §8.3 — Risk + ops → [S]
  • ISO 42001 §8.4 — Data mgmt → [D]
  • ISO 42001 §9.1 — Monitoring → [C]

ISO 27001 Q3 2026 · ISO 42001 Q4 2026 — explicit clause map shipped today.

Customer integration

SCUDO is on by default

Every Claresia tenant gets all five pillars enabled out of the box. Tenants can configure restricted topics, allowed LLM models, egress regions, and SIEM destinations from the cc-059 admin console.

What's automatic
  • Restricted-topics policy seeded with Italian default rules
  • Pre-execution alignment scan on every write-class skill
  • Audit chain emitting to per-tenant Hub stream
  • Article 50 transparency stamp on customer-facing outputs
  • Annex IV technical-doc pack regenerated nightly
  • LLM Gateway routing all calls via eu-south-1 (Milano)
  • Bidirectional Microsoft Presidio PII redaction
What you configure
  • Tenant-specific restricted-topic rules (added on top of defaults)
  • Allowed LLM provider/model pairs per tenant
  • Egress region pinning + non-EU opt-in (requires DPA addendum)
  • SIEM streaming destination (Splunk HEC / Sentinel / Datadog / Elastic)
  • PII policy mode: redact / block / passthrough
  • Telemetry-suppression mode for Statuto dei Lavoratori compliance
  • Hard daily token + cost caps per tenant
  • Roster of approved skill IR hashes
Reference implementation

The TypeScript scaffold backing every pillar lives in cc-aware-governance/lib/governance/:

  • restricted-topics.ts — policy engine
  • alignment-prescan.ts — Haiku-grade alignment check
  • llm-gateway.ts — single egress + redaction + meter
  • eu-ai-act-disclosure.ts — Article 50 stamp
  • audit-chain.ts — SHA-256 chain + verifier
  • tests/governance.test.ts — 22 contract tests, all green
Security review pack

For your CISO + DPO

The pack includes: SCUDO framework spec (PDF), Italian compliance mapping, vs-Glean battlecard, sample audit chain export with the verify-chain CLI, sample Annex IV technical-doc pack, and the latest Italian DPA template.

We email the pack and never share your address. No marketing emails — Trust Center updates are opt-in only.